S7260: Microswimmers on Speed: Simulating Spheroidal Squirmers on GPUs

Transcription

1 S7260: Microswimmers on Speed: Simulating Spheroidal Squirmers on GPUs Elmar Westphal - Forschungszentrum Jülich GmbH

2 Spheroids Spheroid: A volume formed by rotating an ellipse around one of its axes Two kinds: oblate (rotated around its shorter axis), like a pumpkin or teapot prolate (rotated around its longer axis), like an American football 2

3 Squirmers Squirmer is a model for simulating micro swimmers Model origins date back to the 1950s One of today s standard models for self propelled swimmers Can use different means to swim (flagella, arm-like structures etc.), here we simulate the flow caused by surfaces covered with short cilia (filaments) Example: Paramecium bursaria Ein grünes Pantoffeltierchen Frank Fox / license CC BY-SA 3.0 de 3

4 The Simulation Our simulation is split into three parts: Movement of the squirmers and interactions between them Simulation of the liquid Interactions between the squirmers and the liquid 4

5 Simulation of the Squirmers Number of squirmers is low to moderate (up to a few 1000) Simulation of the squirmers and their interactions is sufficiently fast on CPU This may change for future projects 5

6 Simulation of the Fluid We use discrete fluid particles and an algorithm called Multi-Particle Collision Dynamics (MPC) to simulate the fluid MPC inherently conserves energy and momentum of the fluid The phenomena we want to study also require: Conservation of the angular momentum of the fluid particles Adding this roughly doubles the computational effort Walls at one dimension of our system to form a slit This requires additional ghost particles A sufficiently large simulation box for a moderate number of squirmers contains millions of fluid particles 6

7 Simulation of the Fluid We use discrete fluid particles and an algorithm called Multi-Particle Collision Dynamics (MPC) to simulate the fluid MPC inherently conserves energy and momentum of the fluid The phenomena we want to study also require: Conservation of the angular momentum of the fluid particles Adding this roughly doubles the computational effort Walls at one dimension of our system to form a slit This requires additional ghost particles GPU to the Rescue! A sufficiently large simulation box for a moderate number of squirmers contains millions of fluid particles 6

8 Please note that our simulation is in fact a 3-dimensional system. To explain the algorithms, 2D-drawings are used for the sake of simplicity. 7

9 Multi-particle Collision Dynamics (MPC) Fluid particles 8

10 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically One thread per particle, memory-bound 9

11 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically and sorted into cells of a randomly shifted grid One thread per particle, memory-bound 10

12 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically and sorted into cells of a randomly shifted grid For each cell the centre of mass, centre of mass velocity, kinetic energy and angular momentum are computed One thread per particle, atomic-bound, then one thread per cell, memory-bound 11

13 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically and sorted into cells of a randomly shifted grid For each cell the centre of mass, centre of mass velocity, kinetic energy and angular momentum are computed Relative velocities are calculated One thread per particle, memory-bound by random cell-data reads (texture cache) 12

14 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically and sorted into cells of a randomly shifted grid For each cell the centre of mass, centre of mass velocity, kinetic energy and angular momentum are computed Relative velocities are calculated and rotated around a random axis One thread per particle, memory-bound by random cell-data reads (texture cache) 13

15 Multi-particle Collision Dynamics (MPC) Fluid particles are moved ballistically and sorted into cells of a randomly shifted grid For each cell the centre of mass, centre of mass velocity, kinetic energy and angular momentum are computed Relative velocities are calculated and rotated around a random axis Angular momentum and kinetic energy need to be restored several steps using one thread per particle or cell, mostly atomic-bound 14

16 MPC on GPU Is limited by speed of atomic operations and memory bandwidth Implementation uses optimisations as described in some of my earlier GTC talks Reordering of particles to preserve data locality (S2036*) Reducing the number of atomic operations (S5151) *the speed of atomic operations has improved significantly over time, so parts of the implementation described here have been abandoned 15

17 Interactions between Fluid and Squirmers Squirmer surfaces are considered impenetrable for the fluid and are therefore boundaries for the fluid particles Collisions have to be detected Their impact has to be combined and applied to the squirmer and fluid particles accordingly This happens on the GPU (large number of fluid particles), The total impact for each squirmer is passed to and processed by the CPU (low number of squirmers) 16

18 Collisions Between Fluid Particles and Squirmers Fluid particles and Squirmers move during each time-step 17

19 Collisions Between Fluid Particles and Squirmers Fluid particles and Squirmers move during each time-step 17

20 Collisions Between Fluid Particles and Squirmers Fluid particles and Squirmers move during each time-step Squirmer walls are considered impenetrable X 17

21 Collisions Between Fluid Particles and Squirmers Fluid particles and Squirmers move during each time-step Squirmer walls are considered impenetrable Fluid particles entering the squirmer have to be dealt with accordingly 17

22 Collisions Between Fluid Particles and Squirmers Detecting penetration requires rotating particles into the squirmers frame of reference and scaling according to its ratio Checking every fluid particle against every squirmer is too much work, even for a GPU 18

23 Collisions Between Fluid Particles and Squirmers Detecting penetration requires rotating particles into the squirmers frame of reference and scaling according to its ratio Checking every fluid particle against every squirmer is too much work, even for a GPU 18

24 Collisions Between Fluid Particles and Squirmers Detecting penetration requires rotating particles into the squirmers frame of reference and scaling according to its ratio X Checking every fluid particle against every squirmer is too much work, even for a GPU 18

25 Handling Squirmer- Fluid Collisions 19

26 Handling Squirmer- Fluid Collisions The system is divided into cells 19

27 Handling Squirmer- Fluid Collisions The system is divided into cells Regardless of its orientation, a spheroid can not exceed a sphere matching its centre and largest radius 19

28 Handling Squirmer- Fluid Collisions The system is divided into cells Regardless of its orientation, a spheroid can not exceed a sphere matching its centre and largest radius Each squirmer has a list of the cells affected by its surrounding sphere (they may overlap) 19

29 Handling Squirmer- Fluid Collisions The system is divided into cells Regardless of its orientation, a spheroid can not exceed a sphere matching its centre and largest radius Each squirmer has a list of the cells affected by its surrounding sphere (they may overlap) Each cell has a list of the fluid particles it contains 19

30 Handling Squirmer- Fluid Collisions The system is divided into cells Regardless of its orientation, a spheroid can not exceed a sphere matching its centre and largest radius Each squirmer has a list of the cells affected by its surrounding sphere (they may overlap) Each cell has a list of the fluid particles it contains This reduces the number of checks from Nsq x Nfluid to ~Nsq x Vsq* ρ 19

31 Handling Squirmer Surroundings on GPU For each squirmer, a cubical box of cells is defined that contains its surrounding sphere Cells can be conveniently coded into grid- and blockdimensions: x- and y- directions are coded into threadidx z-direction is coded into blockidx.x (limit of 1024 threads per block) Squirmer-index is coded into blockidx.y Affected cells are selected using approximate cell stencils Cells are processed one per thread blockidx.x threadidx.y threadidx.x blockidx.y 20

32 Handling Squirmer Surroundings on GPU Threads from the described grid are used to Select affected cells Check particles from selected cells for collision and Move colliding particles accordingly Sum up collision impact to apply to squirmers Generate random fluid particles inside squirmers to preserve physical properties of the system 21

33 Performance Pitfalls Squirmers and MPC using walls require volumes filled with random particles Their relative impact depends on: Ratio of surface to height of the simulation box (wall algorithm adds an additional cell in one direction) Number and size of squirmers Lack of data locality in squirmer random particles has a severe impact on the performance of the MPC implementation on GPU: Fewer memory accesses can be combined Atomics optimisations depend on data locality 22

34 Computation Time Mostly due to the influence of the spatially unordered random particles inside the squirmers, computation times vary significantly with the number of squirmers (measured on Tesla K80): 4.9 ns per iteration and MPC-particle for large systems with 1 squirmer 13.2 ns per iteration and MPC-particle for large systems with 3692 squirmers Squirmer interactions done on CPU only ~4% of computation time 23

35 General Implementation Heavily templated C++-11 code MPC parts works in 2D, 3D, different precisions and with a variety of options, generating optimised code for different applications and GPU architectures Uses a class template for particle sets to allow mixed precisions and features in a single simulation User managed particle sets can be injected into most steps of the process 24

36 General Implementation Based on a std::vector-like class template using managed memory Replacing the allocator is not enough (members/ operators not defined for device) Easy exchange of data between CPU- and GPU-parts Uses texture caching where applicable Operator templates for CUDA vector-types (float4 etc.) make life a lot easier 25

37 Memory Considerations MPC particles use 52 bytes of memory each MPC cells use 160 bytes of memory each (with angular momentum conservation) Squirmer data uses about 600 bytes of memory Random particles inside squirmers do not count, because they replace particles of the original fluid On recent GPUs, this allows system sizes of up to ~10M cells and ~15K squirmers 26

38 Outlook The next version (currently testing) will definitely feature MPI parallelisation, allowing larger simulations and/or higher speed Hybrid code using template magic to generate CPU or GPU based executables from the same sources (real kernels, not pragma-based) Fewer restraints through the use of even more templates and is prepared for Bringing some (spatial) order to the random particle chaos Bit-true, reproducible calculations 27

39 Example 800 Squirmers in slit, Lx=Lz=300, Ly=7, ~7M fluid particles, 20M timesteps, 2-3 weeks walltime Simulation and rendering courtesy of Mario Theers Further reading: DOI: /C6SM01424K (Paper) Soft Matter, 2016, 12,

40 Example 800 Squirmers in slit, Lx=Lz=300, Ly=7, ~7M fluid particles, 20M timesteps, 2-3 weeks walltime Simulation and rendering courtesy of Mario Theers Further reading: DOI: /C6SM01424K (Paper) Soft Matter, 2016, 12,

41 Thank you for your time Questions? ~100 Squirmers in flow, Lx=300, Ly=Lz=60, ~11M fluid particles, 1M timesteps, 14h walltime Simulation courtesy of Hemalatha Annepu, rendering courtesy of Mario Theers 29

42 Thank you for your time Questions? ~100 Squirmers in flow, Lx=300, Ly=Lz=60, ~11M fluid particles, 1M timesteps, 14h walltime Simulation courtesy of Hemalatha Annepu, rendering courtesy of Mario Theers 29